There is a trend of emerging computing infrastructure aimed at on-demand services, particularly for Internet or other distributed networked computing services. There are basically three categories of on-demand services that are currently available. The first is content delivery, the second is storage, and the third is bandwidth. These services are provided as needed or on-demand, based on a user's needs at any given time. For example, if a first data provider needs greater storage space, an on-demand storage provider simply allocates a greater amount of storage memory to that user, and the first data provider is charged based on the amount of memory space used. If the first data provider no longer needs that amount of memory and deletes information, the on-demand storage provider is then able to re-allocate that memory space to an alternative data provider and the first data provider is charged less because of the reduced storage use.
One of the problems that companies with substantial IT investments face is that it is very difficult for them to predict how much demand they will have for their applications (capacity planning). Therefore, it is extremely difficult for them to determine how large a server farm to deploy which will allow greater user access to their services.
Another problem faced by application or website providers is the continued need for resource capacity to provide adequate service to their users. This is also referred to as the scalability problem. FIG. 1 shows a simplified block diagram representation of the diseconomy of scale result ing in the server infrastructure. What is seen is that application providers are in what is sometimes referred to as a high growth spiral. In the high growth spiral the application provider starts by building a service 52 to gain application users or customers 54. The increase in users results in an increase in the application providers server loads 56. This increased server load causes an increase in response time and often results in the application provider's sites failing or going down, which may result in a loss 60 of users. The application provider must then invest in more resources and infrastructure 62 to reduce response time, improve reliability and keep their users happy 64. This increased response time, and reliability then attracts more users 54, which returns the application provider back to a point where the increased load demands stress or tax the application provider's servers 56, result ing again in a slower response time and a decrease in reliability. Thus, application providers are constantly going around in this high growth spiral.
FIG. 2 shows a graphical representation of the cost per user to increase resource capacity. One of the problems faced by application providers is that the cost of server infrastructure may typically increase faster than the current number of users so that costs are non-linear. This means that as the application provider's server farm gets more complex the cost delta 70 to add enough capacity to service one additional user increases. Thus, the cost 70 of continuing to grow increases dramatically in relation to the cost per user. With most every other business, as the business grows, economies of scale come into effect and the costs per user served actually decreases 72. This is one of the real problems faced by application providers.
Bottlenecks exist in various system resources, such as memory, disk I/O, processors and bandwidth. To scale infrastructure to handle higher levels of load requires increased levels of these resources, which in turn require space, power, management and monitoring systems, as well as people to maintain and operate the systems. As user load increases, so does complexity, leading to costs increasing at a faster rate than volume.
Another problem with providing application processing or services is the amount of capacity that will be needed at start-up, as well as the capacity needs in the future to maintain response time and reliability. These are both start-up costs. It is relatively impossible to predict in advance, with any degree of accuracy, just how successful a site or service is going to be prior to launching and activating the site.
FIG. 3 shows a graphical representation of user capacity demands of an application provider. When an application provider installs a certain number of servers, whatever that number is, the provider has basically created a fixed capacity 74, while demand itself may be unpredictable. Because of the unpredictability of usage demands on servers, that fixed capacity 74 will be either too high 76, and the application provider did not have as many users as anticipated result ing in wasted capacity 76 and wasted capital. Or the fixed capacity 74 was too low 80, and the application provider obtained more users than predicted, result ing in insufficient capacity 80. Thus, if the fixed capacity 74 is too high, the application provider has invested too much capital 76. If the fixed capacity 74 is too low 80, the application provider has users who are dissatisfied because the user does not get the service they need or it takes too long to get responses. This unpredictability is an extremely difficult problem faced by companies providing application processing and services and is particularly severe for those providing services over the Internet simply because of the dynamics of the Internet. The demand is completely unpredictable, and is substantially impossible to plan.
One problem faced by on-line application providers or other users of distributed computing networks is that the network is actually very slow for interactive services as a result of large traverses across the network, because communication signals run into the inherent latency of the network. For example, if an Internet user is in New York, but that New York user want to access a website that is serviced in Los Angeles, the New York user must be routed or hopped all the way across the U.S. Sometimes users will be routed all the way around the world, to get to a specific site. These long distance routings run into large amounts of latency delay. This inherent latency of distributed networks is amplified by the significant increase in the number of interactive services deployed by application and website providers having very active pages or sites. Further, there is a general trend towards customized pages per user. These are sites which are custom created by the server or application for a particular user. These customized sites reduce caching effects to substantially zero. Thus, a customized page, created for that specific user, is generated at the server origin site and routed all the way back across the net to the user adding further inherent delays in the response time. This adds up to a very slow service for more complex interactive services.
In prior art systems, application providers wishing to provide applications have to buy or lease a server, then they must buy or develop the applications that are going to be loaded and run on that server, load the server, and activate the server to provide access to that application. The server is a fully dedicated resource, so that 100% of the time an application is dedicated to a specific server.
Prior art application processing systems require an application provider to route a user to a single central site to allow access to the applications. Every user attempting to access the application is directed to the single central site. Thus, result ing in a bottle neck at the central site. In the prior art single server or single central site, the application provider, however, does maintain access to and control over the application. In some systems where the application provider outsources their server capacity, the application provider must select from a preselected limited number of applications. Further, the application provider no longer has direct control over the application. Any changes desired require the application provider to submit a request to the server provider. Then the server provider must schedule a time at low demands to take the server down to make the changes. This process results in large lag times between the decision to make changes and the implementation of those changes.